Moonshot AI Launches Kimi K3, a 2.8-Trillion-Parameter Open-Weight Model
Chinese lab Moonshot AI released Kimi K3, a ~2.8-trillion-parameter mixture-of-experts model with a 1-million-token context window, built for long-horizon coding and agentic work, with full weights due by July 27.
Moonshot AI released Kimi K3 on July 16, its new flagship model, making it available immediately through Kimi.com, the Kimi app, Kimi Code and the Kimi API, with the full open-weight release to follow by July 27.
What shipped
Kimi K3 is a mixture-of-experts model with roughly 2.8 trillion total parameters, built on a new architecture Moonshot calls Kimi Delta Attention (KDA), a hybrid linear-attention mechanism, combined with a technique the company describes as Attention Residuals. The model supports a 1-million-token context window and native visual understanding, and Moonshot is pitching it specifically at long-horizon coding and agentic work: sustaining multi-step engineering tasks with minimal human supervision, navigating large codebases, and coordinating terminal tools. Two variants launched alongside the base model — K3 Max for chat and agent tasks, and K3 Swarm Max for large-scale parallel processing. On the API, Moonshot is pricing Kimi K3 at $0.30 per million input tokens with caching, $3.00 without, and $15.00 per million output tokens.
Where it stands
Moonshot's own benchmarking, as reported by MarkTechPost, says Kimi K3 delivers frontier-level results across its evaluation suite, consistently outperforming other open-weight models it was tested against, while still trailing the top proprietary systems on the market. The open-weight release is staggered: the model is live for use through Moonshot's hosted products and API now, but the downloadable weights, license and configuration files are not expected until July 27.
Why it matters
Kimi K3 is the latest in a run of large, open-weight releases from Chinese labs this year that trade off narrowly behind the leading proprietary models from OpenAI and Anthropic while remaining freely available to self-host and fine-tune — a dynamic that continues to pressure Western labs' pricing and closed-weight strategies. The model's explicit framing around long-horizon coding and agentic workflows, rather than general chat, also underscores how frontier labs are now optimizing flagship releases specifically for autonomous agent use cases instead of treating agentic capability as a secondary feature.
Sources
AI-assisted reporting, overseen by the AgentsAI team. Spotted an error? Let us know.
More ai news
Sanders and Casar Introduce Bill to Ban 'Artificial Superintelligence' Outright
Sen. Bernie Sanders and Rep. Greg Casar introduced the Ban Artificial Superintelligence Act on September 3, which would permanently prohibit superintelligent AI systems, pause frontier development pending federal safety rules, and impose prison terms and a 'corporate death penalty' for violations.
Anthropic Says Claude Produced the First Machine-Checked Proof of Fermat's Last Theorem
Working largely autonomously for 11 days on the open Prove2Me platform, Claude generated a 13-million-line Lean formalization of Fermat's Last Theorem, which mathematician Kevin Buzzard called an 'extraordinary autoformalization achievement.'
OpenAI Launches GPT-6 Astra, Its First Model Rated 'Critical' Risk for Cyber Offense
OpenAI says Astra is state-of-the-art on coding, computer use and science benchmarks, and is the first model to trip the company's 'critical' cybersecurity threshold — able to find and exploit unknown vulnerabilities without step-by-step human guidance.
Nvidia Agrees to Acquire Hugging Face for $11.9 Billion, Confirming Weeks of Sale Talk
Nvidia will pay $11.9 billion in cash for the open-source AI hub, plus up to $1 billion in retention equity for employees who join, confirming reports from late August that Hugging Face was exploring a sale.